Twelve reasons a take gets thrown away
Discarded takes are not random. Almost all of them belong to one of a dozen recognisable failures, each with a mechanism that explains it and a layer where the repair belongs. Repairing at the wrong layer regenerates the same problem. As of 2026-09-12.
| Symptom | Mechanism | Where the repair belongs |
|---|---|---|
| A character who should be out of shot appears | There is no notion of absence; a name in the text is something to draw | Storyboard |
| A door or object moves with nobody touching it | An action was described by its result rather than by who performs it | Storyboard |
| A hand-off happens but the object stays in both hands | Contact was never written, so the transfer has no moment | Storyboard |
| A prop resets to its opening state mid-scene | Each generation reads the reference afresh and has no memory of the last one | Storyboard |
| An action from the previous segment is performed again | The opening state repeated what had already happened | Storyboard |
| Characters swap sides between shots | The axis was never stated, so each shot picks its own | Storyboard |
| Fingers come out mangled in motion | Hands in fast movement are where generated picture is weakest | Shot design |
| Light appears as a patch on a face | The light was written as where it lands instead of where it comes from | Storyboard |
| The line comes out of the wrong face | The most prominent face receives an unattributed line | Storyboard |
| The reference plate carries debris into every episode | Something from one scene was baked into the plate | Reference art |
| The request is refused outright | Input or output tripped the channel's own review | Channel |
| The same line is spoken twice | The segment ran longer than the dialogue and the spare seconds had no exit | Storyboard |
Inclusion rule. Failures seen often enough to have a named mechanism. A one-off artefact with no reproducible cause is not listed, because nothing can be repaired from it. Order. Alphabetical by symptom.
1Nine of the twelve are storyboard failures
That is the useful finding in the table. The overwhelming majority of discarded takes trace back to something that was ambiguous, absent or wrongly phrased in the written segment, not to a weakness in the picture.
It follows that prompt tinkering at the moment of failure is the least effective response available. The take that failed is gone; what determines the next one is whether the segment now says something different.
2Triage by layer before regenerating anything
Ask which layer owns the problem: the written segment, the reference library, the way the shot was designed, or the channel. Then repair only at that layer and regenerate once.
Regenerating first and diagnosing afterwards is how a production ends up with six takes of the same failure. Each one costs the same as the first, and none of them carries new information.
3Result language is the single most productive thing to remove
Several entries above share one cause: the segment described an outcome instead of what produced it. A door that opens, a light that falls on a cheekbone, a character who is not present. Each is an outcome, and each gets drawn as an outcome.
Rewriting those into causes fixes more than one row at a time. Who pushes the door; where the lamp hangs; who is in the frame, full stop.
4Three of the twelve are not the storyboard failing
Hands in fast motion, debris baked into a reference plate and outright refusals belong elsewhere. The first is a property of what generated picture currently does well, and the repair is to design the shot so the difficult thing is not the subject. The second belongs to the reference library: the plate is remade without what the scene left in it.
The third is the channel exercising its own review. Retrying unchanged achieves nothing; the response is to find which element triggered it and write the beat a different way.
5Keep the count, not the impression
Tally which row each discarded take belongs to across one episode. Most productions find that two rows account for more than half of the waste, and that those two rows correspond to one field of the template being written loosely.
That is a concrete change to make before the next episode, which an impression that the model is unreliable never is.
Craft notes, not a register entry. No line here is attributed to a tool, and none should be read as a claim about one. The sourced material is on the register. Related: One person or a team, What the format demands.
6One pipeline's continuity
Serial work like “Why takes fail” depends on continuity; SceneMixer's guides describe how it keeps it:
SceneMixer generates each character once as a full-body reference sheet — 9:16, plain background, directional key light, identity features written before any technical instruction — and then attaches crops of that sheet to every shot the character appears in.
SceneMixer, AI Character Sheet Generator (read 2026-10-13)
A SceneMixer segment comes out with its sound already in it: each line spoken in that character’s voice, the speaker’s lips moving with the words, and the ambience of the place — rain, traffic, a clock.
SceneMixer, Dialogue and Sound in One Generation (read 2026-10-13)